Deformable Registration of CT Pelvis Images Using Mjolnir Lotta M. Ellingsen and Jerry L. Prince The Johns Hopkins University Image Analysis and Communications Lab 105 Barton Hall, 3400 N. Charles Street, Baltimore, MD 21218 USA E-mail:
[email protected],
[email protected] URL: http://www.iacl.ece.jhu.edu
ABSTRACT Our recently published 3D-3D deformable image registration algorithm, Mjolnir (Ellingsen et al., 2006) was developed for inter-subject registration of MR brain images. Mjolnir is a hybrid registration method, where both anatomical features and image intensities are used to hierarchically align the images. In addition to the hierarchical scheme, the algorithm was implemented in a multi-resolution framework, which both reduces local minima and speeds up the registration process. In this paper an extension to this work is proposed, where Mjolnir has been adapted to register CT images of the pelvis. The main concepts of Mjolnir will be briefly described and the changes made to previous work explained. The algorithm was tested on CT images of the pelvis of 13 different subjects. Results indicate good registration accuracy however, further validation is needed. 1. INTRODUCTION Deformable inter-subject image registration is the process of spatially aligning images of different subjects into a common reference frame so that they can be compared either visually or statistically. During the last few years the need for development of different deformable registration methods has emerged from different clinical applications, such as longitudinal studies and surgical planning [1–7]. The motivation for the extension to our work from MR brain images to CT pelvis images was the need in our research group to create an anatomical atlas for atlas assisted tomography. The proposed registration method was used to register a set of CT images. The resulting displacement fields were then used to align 3D tetrahedral meshes of the subjects to build a statistical anatomical atlas for compensation of missing views in a limited angle cone-beam trajectory [8]. In this paper the basic principles of our previous work on 3D-3D deformable registration of MR brain images [7], upon which this paper builds, will be reviewed. Modifications made to adapt the algorithm to deal with CT pelvis images will be explained and results of the proposed registration method shown. This work was supported 1R21EB003616
in part
by NIH/NIBIB Grant
1-4244-0413-4/06/$20.00 ©2006 IEEE
46
Fig. 1. Pre-processing of the CT data to remove irrelevant elemtents from the image and segment the bone.
2. METHOD 2.1. Pre-processing CT images not only contain objects of interest but also some background elements, like the scanner table. Since our algorithm is based on feature selection and alignment, it was preferred not to have the program spend time aligning features that were irrelevant to our research purpose. Therefore, all such background features were eliminated from the images (see Fig. 1a). Furthermore, since the ultimate objective of our work was to use the results to generate a pelvic bone atlas, the strong edge feature between the background and the soft tissue of the patient was removed. This also led to better registration results of the bone because of the challenge of allowing large displacements in the soft tissue while maintaining more rigidity in the bone structure due to much smaller variability in bone tissue size compared to soft tissue size (see Fig. 2). Fi-
NORSIG 2006
9 × 1 vector of geometric moment invariants (GMIs). The GMIs are formulated from the zero-order and second-order 3D regular moments in the following way:
Fig. 2. High variability in soft tissue compared to bone tissue between different subjects.
I1
= M0,0,0
I2
= M2,0,0 + M0,2,0 + M0,0,2
I3
= M2,0,0 × M0,2,0 + M2,0,0 × M0,0,2 2 2 2 + M0,2,0 × M0,0,2 − M1,0,1 − M1,1,0 − M0,1,1 ,
where Mp,q,r = ZZZ
xp1 xq2 xr3 fm (x1 , x2 , x3 ) dx1 dx2 dx3
(x1 )2 +(x2 )2 +(x3 )2
